Migrate cspell config to per-service files ***NO_CI***

[maorleger]

What

Adopt the latest Azure SDK spellcheck guidance
(spellcheck.md):

• Move service-scoped overrides out of the central .vscode/cspell.json
  into per-service sdk/<service>/cspell.yaml files that import the central
  config (the import is required).
• Convert legacy per-service cspell.json → cspell.yaml, add the missing
  import to existing per-service cspell.yaml files, and consolidate each
  service down to a single sdk/<service>/cspell.yaml.
• Alphabetize word lists (case-insensitive) and dedupe.

Cross-cutting words stay in the central config (they have no reliable
per-service attribution).

Why

Keeping service words in per-service files reduces merge conflicts in the hot
central file and makes the lists easier to read — exactly what the guidance now
recommends.

Safety / validation

• cspell uses the nearest config for each file, so consolidating package
  configs up to the service directory preserves coverage (verified empirically
  with cspell 10, the pinned version).
• Conservation check: the full case-insensitive word union is identical
  before and after — 0 words lost.
• Spot-checked affected files (e.g. voicelive/keyvault review docs, storage &
  cosmos sources) still report 0 issues.

CI

Config-only change across many packages, so the commit is tagged ***NO_CI***
to avoid a repo-wide build storm. Opened as a draft.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
7 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[maorleger]

Reference: this migration follows the Azure SDK spellcheck guidance — Migrating from .vscode/cspell.json to sdk/<service>/cspell.yaml.

[maorleger]

Updated to fix the spell-check failure seen in the earlier CI run
(build 6573147).

Root cause: the initial version kept service words inside per-file overrides
blocks. When CI spell-checks the changed cspell.yaml file itself, top-level
words self-cover it but words nested in overrides do not, so tokens like
BYOK/coreai were flagged as unknown in sdk/agentserver/cspell.yaml.

Fix: follow the guidance literally — one sdk/<service>/cspell.yaml per
service with all words flattened to top-level words (no overrides). The file
now self-covers, and a single config per service keeps words scoped to that
service. Verified locally with cspell 10: every generated cspell.yaml passes a
self-lint, no words were lost vs. the previous config, and affected source files
still pass.